How US Sanctions Built China's AI Chip Industry: The Cambricon Miracle and the SMIC Bottleneck

*China's domestic AI chip ecosystem has grown from a pipe dream into a trillion-yuan reality. The biggest catalyst wasn't government subsidy — it was American export controls. (Image: Unsplash)*

On the morning of August 8, 2026, Cambricon Technologies released its first-half earnings report, and the numbers told a story that would have seemed absurd just two years earlier. Revenue: ¥5.996 billion, up 108% year-over-year. Net profit: ¥2.311 billion, up 123%. Gross margin: robust and expanding. And perhaps most strikingly, the company's market capitalization had crossed ¥1 trillion in June — making it the first company on Shanghai's STAR Market to achieve that milestone.

For a company that had burned through cash for nine consecutive years, this wasn't just a turnaround. It was a vindication of an entire industrial strategy. But the most remarkable part wasn't in the earnings report itself. It was in what the report didn't say: Cambricon's meteoric rise was made possible by the very policy designed to destroy it.

The US government's export controls on advanced AI chips — first imposed in 2022, tightened repeatedly through 2023 and 2024, and expanded again in early 2025 — were intended to slow China's artificial intelligence development by cutting off access to NVIDIA's most powerful GPUs. Instead, they created the market conditions that transformed China's domestic chip industry from a collection of loss-making startups into a profit-generating, capital-attracting, ecosystem-building powerhouse.

This is the great semiconductor irony of 2026: the sanctions worked. They worked so well they backfired.

The Conventional Wisdom: Sanctions as Suffocation

The dominant narrative about US-China tech competition, repeated endlessly in Washington think tanks, congressional hearings, and Wall Street research notes, goes something like this: without access to cutting-edge semiconductor manufacturing equipment and NVIDIA's industry-standard GPUs, China's AI industry is stuck. The country can design chips but can't manufacture them at competitive yields. Its domestic alternatives are years behind. The talent gap is unbridgeable. Eventually, the pressure will force Beijing to the negotiating table.

This narrative isn't entirely wrong — it's just dangerously incomplete. Yes, SMIC's most advanced node (N+3, roughly equivalent to 5nm-class) trails TSMC's 2nm by a full generation. Yes, Chinese AI chips generally consume more power and deliver less performance per watt than their American counterparts. Yes, the software ecosystem around CUDA remains a formidable moat.

But the narrative makes a critical error: it assumes that cutting off supply to a market of 1.4 billion people and the world's second-largest economy will cause that market to simply wither. In reality, when you deny a major economy access to essential technology, you don't eliminate demand. You redirect it. And if that economy has sufficient capital, engineering talent, and government commitment, redirected demand becomes the foundation of a domestic industry.

That's exactly what happened in China.

The Evidence: Numbers That Defy the Narrative

Cambricon's financial trajectory over the past 18 months reads like a case study in how to build an industry through regulatory arbitrage. The company went from chronic losses to its first full-year profit in 2025 — ¥2.059 billion in net profit on ¥6.497 billion in revenue, a staggering 453% revenue increase year-over-year. The first half of 2026 built on that momentum with another doubling of revenue and profit.

*Data sources: Cambricon H1 2026 earnings report, STAR Market filings*

But Cambricon is only the most visible success story. The entire ecosystem is booming:

Hygon, Cambricon's closest domestic rival, reported H1 2026 revenue of ¥8.5–9.3 billion (up 55–70% YoY) and net profit of ¥1.7–1.83 billion (up 41–52%). Excluding stock-based compensation, underlying profit growth approached 85%.

Moore Threads, a GPU-focused startup that listed on STAR Market in December 2025, posted H1 2026 revenue of ¥1.736 billion — already exceeding its entire 2025 full-year revenue of ¥1.506 billion. Its net loss narrowed 95.7% to just ¥11.56 million, effectively reaching breakeven. The company is now planning a dual Hong Kong listing.

Biren, founded by NVIDIA alumni, raised HK$5.58 billion in its January 2026 Hong Kong IPO. MetaX and Iluvatar CoreX have also gone public, and Enflame — with Tencent as a 20.3% shareholder — received IPO approval in July 2026 to raise ¥6 billion.

The aggregate picture is startling. According to TrendForce and industry analyses, domestic Chinese AI chip solutions are on track to capture nearly 90% of China's high-end AI chip market in 2026 — up from negligible market share just three years ago. The China market alone consumed approximately 4 million domestic AI chips in 2025, with projections of 5 million units in 2026.

These aren't subsidy-dependent zombies. These are companies with real customers, real revenue, and in Cambricon's case, real profits at scale.

The Real Story: How Sanctions Created a Captive Market

To understand why sanctions backfired, you have to understand what they actually did. The US didn't ban all semiconductor sales to China — it banned the most advanced AI chips, specifically NVIDIA's A100, H100, and subsequent datacenter GPUs. This created a peculiar market dynamic: Chinese companies could still buy lower-tier NVIDIA chips, but for training large AI models — the core activity driving demand — those chips were either unavailable or legally risky to acquire.

The immediate result was a supply vacuum at the exact moment Chinese AI demand was exploding. ByteDance's Doubao assistant surged past 180 million monthly active users. Alibaba's Qwen models became the most-downloaded open-source AI family globally. DeepSeek proved that Chinese labs could build frontier-class models with domestic chips. Every major Chinese tech company suddenly needed AI compute at scale, and their traditional supplier was legally barred from meeting that need.

Into this vacuum stepped the domestic chip industry. But here's the crucial part: the domestic industry didn't just get orders. It got the right kind of orders.

ByteDance, Cambricon's largest customer, reportedly placed pre-orders for 200,000 Cambricon chips — the kind of volume commitment that allows a chip designer to achieve manufacturing scale, amortize R&D costs, and negotiate better terms with foundries. Without US sanctions, ByteDance would have bought NVIDIA. With sanctions, it bought Cambricon. And because the order was large enough to justify dedicated production lines, Cambricon could invest in yield improvement, software optimization, and next-generation architecture.

The same dynamic played out across the ecosystem. The Chinese government, which had been promoting domestic chip adoption for years with limited success, suddenly found that the market was doing its job for it. In May 2026, nine domestic AI chips were certified for government procurement — not because of a new policy push, but because they had finally reached performance thresholds where buying domestic was no longer a political statement. It was a rational business decision.

The SMIC Bottleneck: The Constraint That Keeps the Dream Grounded

If the story ended here, it would be a clean tale of unintended consequences and industrial policy success. But reality is messier. For all the domestic chip designers' success, they share a single, critical dependency: SMIC.

Semiconductor Manufacturing International Corporation is China's most advanced chip foundry, and it is the production backbone for virtually every domestic AI chip designer. SMIC's Q2 2026 results tell their own story of demand outstripping supply: $3.01 billion in revenue (up 36% YoY), gross margin expanding to 25.3% (from 20.4% a year earlier), and management explicitly citing the need to "accelerate ramp-up of new capacity to relieve industry-wide supply constraints."

The problem is fundamental. SMIC's most advanced production node, N+3 (roughly 5nm-class), only entered volume production in mid-2026. Its 7nm-equivalent capacity (N+2) is estimated at 45,000–60,000 wafers per month — a fraction of what TSMC produces. And yields on advanced nodes remain challenging. Industry estimates suggest N+2 yields in the 20% range — meaning four out of five chips fail quality control.

*Data compiled from company reports and industry estimates*

Do the math: at 60,000 wafers per month, ~20% yield, and roughly 500 chips per wafer, SMIC can produce approximately 2.6 million advanced AI chips per year. China needs 5 million in 2026. The gap is structural, not cyclical.

This creates a bizarre situation. Cambricon, Moore Threads, Hygon, and their peers are designing increasingly sophisticated chips that are, by many accounts, competitive with mid-tier NVIDIA products on raw performance. But they can't manufacture enough of them. The bottleneck isn't American design superiority anymore. It's Chinese manufacturing capacity.

SMIC is racing to expand. The company plans to flexibly deploy existing fabs and accelerate new capacity. But building a leading-edge semiconductor fabrication plant costs $15–20 billion and takes 3–4 years. SMIC's 2025 revenue of $9.3 billion — impressive as it is — doesn't generate the cash flow to fund multiple advanced fabs simultaneously. And every yield point improvement on N+2 or N+3 requires expensive process development that SMIC must fund while simultaneously building the next node.

The result is a two-speed industry. The design side is thriving. The manufacturing side is straining. And the gap between them is where China's AI chip story will be won or lost.

The Competitive Landscape: From NVIDIA-or-Nothing to a Crowded Field

Perhaps the most underappreciated effect of US sanctions has been the diversification of China's AI chip supply chain. Three years ago, the market was binary: NVIDIA or nothing. Today, Chinese AI infrastructure buyers have a menu of options.

*Data compiled from company filings, earnings reports, and industry estimates*

This landscape reveals something important: the Chinese AI chip industry isn't just one company catching up. It's an entire ecosystem maturing simultaneously. Each player occupies a slightly different niche. Cambricon dominates cloud inference with its MLU series. Hygon specializes in datacenter accelerators compatible with CUDA-like programming models. Moore Threads pursues general-purpose GPU computing. Huawei's Ascend line targets both training and inference with the deepest software stack.

The diversity matters because it creates resilience. If one company's architecture proves unsuited to next-generation AI models, others can fill the gap. If one foundry partnership falters, alternatives exist. The US sanctions, by forcing fragmentation, may have inadvertently created a more robust domestic supply chain than any single-company strategy could have achieved.

Software: The Silent Battleground

Hardware is only half the story. The real moat in AI chips isn't the silicon — it's the software stack that makes the silicon usable. NVIDIA's dominance wasn't built on GPU architecture alone; it was built on CUDA, the programming framework that became the industry standard for AI development.

Chinese chip companies understand this. Cambricon has invested heavily in its BANG C programming language and Neuware software platform, which now supports major Chinese open-source models including DeepSeek, Qwen, GLM, Kimi, and MiniMax. The company's H1 2026 report notes explicitly that it "expanded support for major Chinese open-source AI models" — a recognition that hardware without software is just expensive sand.

But the gap remains real. A developer trained on CUDA can't seamlessly switch to BANG C or Huawei's CANN framework. The ecosystem of pre-built models, optimization libraries, and community knowledge around NVIDIA is decades in the making. Chinese alternatives are improving rapidly — particularly for inference workloads, where the programming model is simpler — but training large models on domestic chips still requires significant engineering investment that many companies would prefer to avoid.

This creates a paradox. The domestic chip industry has the orders, the revenue, and the manufacturing pipeline. But the software ecosystem is still playing catch-up. And until a Chinese programming framework achieves something close to CUDA's ubiquity, the industry will remain dependent on NVIDIA for the most demanding training workloads — even if those NVIDIA chips have to be acquired through gray-market channels or used in overseas data centers.

The Investment Tsunami: Capital Flows Where Returns Exist

One of the most striking features of China's AI chip boom is how it has transformed capital markets. The STAR Market — Shanghai's Nasdaq-style exchange for technology companies — has become a listing destination of choice for chip designers. In the past year alone, Moore Threads, MetaX, and Cambricon's continued rally have demonstrated that Chinese investors are willing to pay premium valuations for domestic semiconductor plays.

Cambricon's ¥1 trillion market cap milestone in June 2026 wasn't just a symbolic achievement. It signaled that the market believes domestic AI chips are a permanent feature of China's technology landscape, not a temporary workaround until sanctions are lifted. When a loss-making startup can become one of China's most valuable technology companies in under a decade, the investment calculus for the entire sector changes.

Venture capital has followed. Chinese government guidance funds, state-backed investment vehicles, and private VC firms have poured billions into chip design startups. The 2025 total of ¥107 billion in R&D spending by China's A-share listed chip companies — a record — reflects not just government subsidy but genuine investor enthusiasm for an industry that is finally generating returns.

The dual-listing trend — STAR Market plus Hong Kong — is particularly telling. Moore Threads, MetaX, and others are pursuing Hong Kong IPOs not because they need the capital (though they do), but because Hong Kong provides access to international investors who might otherwise be wary of mainland listings. It's a recognition that China's AI chip industry is becoming globally relevant, even if its products remain largely domestic.

Implications: Who Wins, Who Loses

The great semiconductor irony of 2026 produces clear winners and losers.

Winners:

- Chinese AI chip designers — Cambricon, Hygon, Moore Threads, and the broader ecosystem have gone from science projects to profitable businesses with real market power.

- SMIC — As the monopoly domestic foundry for advanced chips, SMIC has pricing power it never had before. Every yield improvement drops directly to the bottom line.

- Chinese AI labs — Companies like ByteDance, Alibaba, and DeepSeek now have access to domestic compute at scale, insulating them from future supply disruptions.

- The Chinese government — Achieved, through American policy rather than its own industrial planning, the domestic chip adoption it had been pursuing for a decade.

Losers:

- NVIDIA — Lost what was once its second-largest market. Even the company's specially-designed China-compliant chips (the H20 and its successors) face growing competition from domestic alternatives that are improving faster than expected.

- US semiconductor equipment makers — Applied Materials, Lam Research, and KLA have lost billions in potential sales as China redirects its capital to domestic equipment suppliers.

- The US policy establishment — Export controls achieved the opposite of their intended effect. Rather than slowing China's AI development, they accelerated the creation of a domestic chip industry that will eventually compete globally.

Uncertain:

- TSMC — As the world's most advanced foundry, TSMC benefits from the global chip boom. But if China's domestic industry matures to the point where it can design chips competitive with TSMC's American customers, the geopolitical dynamics of semiconductor manufacturing could shift dramatically.

The Future: From Domestic to Global

The most important question about China's AI chip industry isn't whether it can survive without NVIDIA. It already has. The question is whether it can expand beyond China's borders.

Currently, the answer is: not yet. Cambricon's revenue is 99.98% from its cloud product line, and that cloud product line is almost entirely domestic. Huawei's Ascend chips, despite their technical sophistication, face export control restrictions that limit their international deployment. The software ecosystem remains too China-centric for global developers to adopt easily.

But history suggests that domestic scale precedes global expansion. Samsung's semiconductor business was built on supplying Korea's electronics industry before it became a global powerhouse. TSMC's early customers were Taiwanese PC motherboard makers before it became the world's foundry. China's AI chip companies are following the same playbook: dominate the home market, build scale, improve yields and software, then compete internationally.

The timeline is uncertain. SMIC's capacity constraints mean that even domestic demand may outstrip supply for the next 2–3 years. Software ecosystems take years to mature. And geopolitical headwinds — particularly the threat of broader sanctions on Chinese chip exports — could delay global expansion indefinitely.

But the direction is clear. An industry that didn't exist as a commercial proposition three years ago is now generating tens of billions of yuan in revenue, attracting trillions in market capitalization, and designing chips that are genuinely competitive for an increasing share of AI workloads. The sanctions created it. The market is sustaining it. And the technology is improving fast enough that counting it out would be a mistake.

Social Voices: What Analysts and Developers Are Saying

X (Twitter) — "Cambricon hitting ¥1T market cap is the ultimate proof that US sanctions are the best industrial policy China never had to pay for. Nine years of losses, then one export control order and suddenly they're profitable. You can't make this up." — @semiconductor_watcher

Zhihu — "寒武纪的财报确实 impressive，但别忽视了背后的 SMIC 产能瓶颈。设计再好，造不出来也是白搭。20% 的良率意味着每片晶圆只有 100 颗能用芯片，这个成本结构能不能持续盈利还是未知数。" / "Cambricon's earnings are indeed impressive, but don't ignore the SMIC capacity bottleneck behind it. Great designs are useless if you can't manufacture them. A 20% yield means only ~100 usable chips per wafer — whether this cost structure can sustain profitability is still unknown."

Weibo — "美国打芯片牌打了四年，打出来一个寒武纪万亿市值、海光利润翻倍、摩尔线程差点盈利。这算哪门子的制裁？这分明是在帮中国做产业升级。" / "America played the chip card for four years and produced a Cambricon with trillion-yuan market cap, Hygon with doubled profits, and Moore Threads nearly breaking even. What kind of sanctions are these? This is clearly helping China's industrial upgrading."

Douban — "有点担心这个繁荣是虚假的。一旦美国放松制裁，NVIDIA 重新进来，这些国产芯片公司还能不能保持竞争力？现在的利润是不是建立在‘没有 NVIDIA’的温室环境里？" / "I'm a bit worried this prosperity is artificial. If America eases sanctions and NVIDIA re-enters, can these domestic chip companies remain competitive? Are current profits built in a greenhouse environment where NVIDIA doesn't exist?"

Hacker News — "The yield numbers are the real story. 20% at 7nm means SMIC is still years behind TSMC, but they're learning fast. The fact that Cambricon is profitable *despite* those yields means either their chips are selling at massive premiums or their designs are so efficient they work even on flawed silicon. Either way, it's impressive."

Xiaohongshu — "在字节工作，我们内部确实在大量采购寒武纪的芯片。不是政治任务，是性价比真的不错，而且供应链稳定。DeepSeek 模型在 MLU 上跑得很顺。" / "Working at ByteDance, we're indeed purchasing Cambricon chips in large quantities. Not a political task — the cost-performance is genuinely good, and the supply chain is stable. DeepSeek models run smoothly on MLU."

*This article was published on August 14, 2026. For more analysis of China's AI landscape, see our coverage of Huawei's Ascend AI ecosystem, Tencent's AI turnaround, and China's AI model export boom.*